Classification under input uncertainty with support vector machines

Uncertainty can exist in any measurement of data describing the real world. Many machine learning approaches attempt to model any uncertainty in the form of additive noise on the target, which can be effective for simple models. However, for more complex models, and where a richer description of anisotropic uncertainty is available, these approaches can suffer. The principal focus of this thesis is the development of advanced classification approaches that can incorporate the known input uncertainties into support vector machines (SVMs), which can accommodate isotropic uncertain information in the classification. This new method is termed as uncertainty support vector classification (USVC). Kernel functions can be used as well through the derivation of a novel kernelisation formulation to generalise this proposed technique to non-linear models and the resulting optimisation problem is a second order cone program (SOCP) with a unique solution. Based on the statistical models on the input uncertainty, Bi and Zhang (2005) developed total support vector classification (TSVC), which has a similar geometric interpretation and optimisation formulation to USVC, but chooses much lower probabilities that the corresponding original inputs are going to be correctly classified by the optimal solution than USVC. Adaptive uncertainty support vector classification (AUSVC) is then developed based on the combination of TSVC and USVC, in which the probabilities of the original inputs being correctly classified are adaptively adjusted in accordance with the corresponding uncertain inputs. Inheriting the advantages from AUSVC and the minimax probability machine (MPM), minimax probability support vector classification (MPSVC) is developed to maximise the probabilities of the original inputs being correctly classified. Statistical tests are used to evaluate the experimental results of different approaches. Experiments illustrate that AUSVC and MPSVC are suitable for classifying the observed uncertain inputs and recovering the true target function respectively since the contamination is normally unknown for the learner.

Name dc:type.qualificationname

Ph.D.

Level dc:type.qualificationlevel

doctoral

Grantor dc:publisher.institution

University of Southampton

Year dc:date.issued

2009

Author dc:creator

• Yang, Jianqiang

Advisor dc:contributor.advisor

• Gunn, S.